Stable Audio 3 Inference Methods

# Stable Audio 3 Inference Methods An overview of the different inference modes. The python interface is shown, but these controls are the same as for the gradio interface > New to diffusion/Flow Matching models? See [Model Overview](../guides/model-overview.md) > for a conceptual overview before diving in. | Model | Type | |---|---| | `medium` | Post-trained | | `small-music` | Post-trained | | `small-sfx` | Post-trained | | `medium-base` | Base | | `small-music-base` | Base | | `small-sfx-base` | Base | > **Note:** `medium` and `medium-base` require a CUDA GPU with Flash Attention support due to using SAME-L as their autoencoder. - **`prompt`** — Text description of the audio to generate. For help crafting good prompts, see [Prompt Guide](../guides/prompting.md) - **`duration`** — Duration of the generated audio in seconds (default: `120`). - **`steps`** — Number of sampling steps (default: `8`). For even faster inference, reduce this number at some cost to quality. However, going higher than 8 doesn't necessarily increase quality (unless using a '-base' model, where you should use something like 50) - **`seed`** - Random seed for reproducible outputs if needed. Use -1 to select a random seed (default) or select your favorite number for deterministic results. - **`batch_size`** - Generate multiple at once, useful is you have a GPU and want to get a lot of variations. The max is limited by your GPU's VRAM.